Predicting Human Emotions through Time and Frequency Feature Extraction using EEG Emotional Database Analysis

Emotion recognition is a technique to identify the emotion of an individual by using facial expressions, behavior, and neuroimaging techniques. Electroencephalography (EEG) is a non-invasive technique mainly used to recognize emotions. To study variations in human emotions, the authors created an EEG dataset, named EEG Emotional Database Analysis (EEDA), by conducting a study with the help of an EEG technician on 10 healthy participants. The participants were asked to watch video clips representing five emotions with a 60-second gap between clips. The emotional responses were collected simultaneously through EEG across five trials, utilizing a 32-channel system. The EEG data were preprocessed to remove artifacts using a band-pass filter and Independent Component Analysis (ICA). Subsequently, signal features were extracted in three domains: time, frequency, and time-frequency. The data were segregated into five classes: Relaxed-Happy-Excitement (RHE), Relaxed-Fear-Sad (RFS), Relaxed-Happy-Fear (RHF), Relaxed-Excitement-Sad (RES), and Relaxed-Happy-Sad (RHS) for training and testing purposes. After feature extraction, feature selection was performed using Principal Component Analysis (PCA), mutual information, Recursive Feature Elimination (RFE), and Least Absolute Shrinkage and Selection Operator (LASSO) regularization to calculate the accuracy with the Bidirectional Long Short-Term Memory (BiLSTM) model. The highest classification accuracies obtained on the EEDA dataset were 84.27% for the RHS class in the time domain, 61.81% for the RES class in the frequency domain, and 84.37% in the time-frequency domain for the RHS and RHF classes.